Project chat

Project chat is a read-only assistant embedded in the project and workspace shells. It answers questions about what exists — the project, its members, its status — without proposing changes. It is deliberately *not* the Plan chat, which authors Work Orders.

Two scopes exist, distinguished by mode:

workspace_level  -> the workspace home assistant
project_level    -> the assistant in the project shell

What it can do

Project chat is informational. In workspace_level mode its tools are limited to safe, read-only queries — workspace info, member lists, the active model, the project list. It will tell you how the workspace is doing or who is in it; it will not create, approve, or mutate anything. Sessions are shared across a project's members, so the conversation history is a shared artifact.

Why it is separate from Plan

Keeping the two apart is a safety boundary. Plan proposes structured Work Orders and can change the plan; project chat cannot. A member asking "what's the status of the login work?" should never risk creating a Work Order by accident. The two run different tool allowlists for exactly this reason.

Under the hood

Both modes are the same Chat entity with a different mode and tool allowlist, backed by a conversation history that is loaded on open and appended on each turn. The tools it calls are a subset of the MCP tool catalog.

Next: Workspaces & projects.